Classification of Respiratory Abnormalities using Deep Learning Techniques

Abstract

Chronic Obstructive Pulmonary disease (COPD) is the most prevailing and progressive respiratory disorder. Measurement of respiratory mechanics helps in monitoring pulmonary disease. Exacerbations of COPD cause the disease to become more aggressive, physical functions to deteriorate, and results in declined life quality. Prevention of exacerbations by providing optimizing treatment becomes mandatory. The most frequently used clinical test to measure the volume of lung is spirometry. It depends on the maximal effort of the patient for perfect diagnosis of pulmonary disorder. An attempt has been made in this work to detect pulmonary abnormality using flow-volume spirometry. A portable Spirolab II spirometer was used to collect respiratory data from 1030 subjects (N = 1030) using a standard acquisition Procedure. Totally 1030 subjects with16 features were consider for investigation purpose. The categorization of normal and abnormal groups of spirometric data were performed using deep learning techniques. Deep architectures have drawn a lot of interest from a variety of domains due to their representational power. It is one of the most powerful architecture of unsupervised learning, which acquires features through empirical observation from data. The effort of the architecture to model high level abstractions in data allows to perform automated analysis of raw physiological data with minimal human intervention. The proposed work uses Restricted Boltzmann Machines with variable learning rate (DRBM) and Deep Belief Network (DBN) for classification of spirometric data. This approach was demonstrated to have the capability to classify the spirometric patterns in to normal and obstructive classes using DRBM and normal restrictive and obstructive classes using DBN. Further validation is carried out using k-fold cross validation. This show that, with raw spirometric data DRBM and DBNs allow us to achieve an average accuracy of 98% for two class and 85.58 % for three class classification respectively. Thus this analysis appears to provide better clinical significance.